`before_replace` triggers inconsistency when updating a primary key

[sergepetrenko]

Tarantool version: 2.11.0-entrypoint-358-g9965e3fe4 (current master at the time of writing)

I've noticed some strange behaviour with our before_replace triggers:

1. A usual insert of a duplicate key to a unique index leads to an ER_TUPLE_FOUND error. However, when the space has a before_replace trigger, the same insert completes just fine:

tarantool> box.cfg{}
---
...

tarantool> box.once('bootstrap', function()
         > box.schema.space.create('test')
         > box.space.test:create_index('pk')
         > end)
---
...

tarantool> box.space.test:insert{1}
---
- [1]
...

tarantool> box.space.test:insert{1} -- duplicate 
---
- error: Duplicate key exists in unique index "pk" in space "test" with old tuple
    - [1] and new tuple - [1]
...

tarantool> box.space.test:before_replace(function(old, new) return new:update{{'=', 1, new[1]}} end)
---
- 'function: 0x0111209ec0'
...

tarantool> box.space.test:insert{1} -- same duplicate, but succeeds with a no-op before_replace trigger!
---
- [1]
...

2. Usually it's not allowed to change tuple's fields indexed by primary key in before_replace triggers. But when inserting a tuple not present in space this works just fine, and also allows duplicates like in example 1:

tarantool> box.cfg{}
---
...

tarantool> box.once('bootstrap', function()
         > box.schema.space.create('test')
         > box.space.test:create_index('pk')
         > end)
---
...

tarantool> box.space.test:insert{1}
---
- [1]
...

tarantool> box.space.test:select({}, {fullscan=true})
---
- - [1]
...

tarantool> box.space.test:before_replace(function(old, new) return new:update{{'=', 1, 2}} end)
---
- 'function: 0x0110d1da30'
...

tarantool> box.space.test:insert{1} -- error as expected
---
- error: Attempt to modify a tuple field which is part of primary index in space 'test'
...

tarantool> box.space.test:insert{3} -- no error for a new tuple!
---
- [2]
...

tarantool> box.space.test:insert{4} -- and even duplicates!
---
- [2]
...

So, first of all, seems like duplicate keys should be checked after before_replace triggers are executed, not earlier. And, secondly, we should either disallow updating fields indexed by primary key completely or allow such updates and handle them correctly in all cases.

[olegrok]

tarantool> box.space.test:insert{1} -- same duplicate, but succeeds with a no-op before_replace trigger!
---
- [1]
...

Probably it's known feature https://github.com/tarantool/tarantool/issues/4503#issuecomment-557903138